1. Case Study 1: Data Replication for LIGO Scott Koranda Ann Chervenak

2. eSS 2007Service-Oriented Science: Globus Software in Action2 Laser Interferometer Gravitational Wave Observatory l Goal: Observe gravitational waves predicted by theory l Three physical detectors in two locations (plus GEO detector in Germany) l 10+ data centers for data analysis l Collaborators in ~40 institutions on at least three continents

3. eSS 2007Service-Oriented Science: Globus Software in Action3 LIGO by the Numbers l LIGO records thousands of channels, generating approx. 1 TB per day of data during a detection run u Data is published and data centers subscribe to portions that local users want for analysis or for local storage l Data analysis results in derived data u Currently ~30% of all LIGO data u This also is published and replicated l Over 30 million files in LDR network (April 2005) * l 6+ million unique logical files l 30+ million physical copies of those files * Scott Koranda, UW-Milwaukee, April 2005

4. eSS 2007Service-Oriented Science: Globus Software in Action4 The Challenge Replicate 1 TB/day of data to 10+ international sites u Publish/subscribe (pull) model u Provide scientists with the means to specify and discover data based on application criteria (metadata) u Provide scientists with the means to locate copies of data

5. eSS 2007Service-Oriented Science: Globus Software in Action5 Issues - Technical l Efficiency u Avoid unused bandwidth while data transfers are taking place, esp. on high- bandwidth links (10+ Gbps) u Avoid idle time on the network between transfers

6. eSS 2007Service-Oriented Science: Globus Software in Action6 Issues - Social (1) l Workflow u The publish/subscribe model matches how scientists think about their use of the data u Use of metadata is critical l Security u Authenticate endpoints of data transfers to prevent unauthorized data manipulation l Heterogeneity u Cant tell data centers what storage systems to use u Cant get everyone to do accounts the same way

7. eSS 2007Service-Oriented Science: Globus Software in Action7 Issues - Social (2) l Maintenance u LIGO is focused on science, not IT u Ideally, they would not have to build or maintain data management software u GOAL: If software must be produced, keep it simple and do it in a way that non-CS people can understand u GOAL: Produce a solution that others will adopt and maintain for them

8. eSS 2007Service-Oriented Science: Globus Software in Action8 GridFTP l A high-performance, secure data transfer service optimized for high- bandwidth wide-area networks u FTP with extensions u Uses basic Grid security (control and data channels) u Multiple data channels for parallel transfers u Partial file transfers u Third-party (direct server- to-server) transfers l GGF recommendation GFD.20 Basic Transfer One control channel, several parallel data channels Third-party Transfer Control channels to each server, several parallel data channels between servers

9. eSS 2007Service-Oriented Science: Globus Software in Action9 Striped GridFTP l GridFTP supports a striped (multi-node) configuration u Establish control channel with one node u Coordinate data channels on multiple nodes u Allows use of many NICs in a single transfer l Requires shared/parallel filesystem on all nodes u On high-performance WANs, aggregate performance is limited by filesystem data rates

10. eSS 2007Service-Oriented Science: Globus Software in Action10 globus-url-copy l Command-line client for GridFTP servers u Text interface u No interactive shell (single command per invocation) l Many features u Grid security, including data channel(s) u HTTP, FTP, GridFTP u Server-to-server transfers u Subdirectory transfers and lists of transfers u Multiple parallel data channels u TCP tuning parameters u Retry parameters u Transfer status output

11. eSS 2007Service-Oriented Science: Globus Software in Action11 RFT - File Transfer Queuing l A WSRF service for queuing file transfer requests u Server-to-server transfers u Checkpointing for restarts u Database back-end for failovers l Allows clients to request transfers and then disappear u No need to manage the transfer u Status monitoring available if desired

12. eSS 2007Service-Oriented Science: Globus Software in Action12 RLS - Replica Location Service l A distributed system for tracking replicated data u Consistent local state maintained in Local Replica Catalogs (LRCs) u Collective state with relaxed consistency maintained in Replica Location Indices (RLIs) l Performance features u Soft state maintenance of RLI state u Compression of state updates u Membership and partitioning information maintenance Simple Hierarchy The most basic deployment of RLS Fully Connected High availability of the data at all sites Tiered Hierarchy For very large systems and/or very Large collections

13. eSS 2007Service-Oriented Science: Globus Software in Action13 pyGlobus * l High-level, object- oriented interface in Python to GT4 Pre- WS APIs. u GSI security u GridFTP u GRAM u XIO u GASS u MyProxy u RLS * pyGlobus contributed to GT4 by the Distributed Computing Department of LBNL l Also includes tools and services u GridFTP server u GridFTP GUI client u Other GT4 clients

14. eSS 2007Service-Oriented Science: Globus Software in Action14 Lightweight Data Replicator * Ties together 3 basic Grid services: 1. Metadata Service info about files such as size, md5, GPS time,... sets of interesting metadata propagate answers question What files or data are available? 2. Globus Replica Location Service (RLS) catalog service maps filenames to URLs also maps filenames to sites answers question Where are the files? 3. GridFTP Service server and customized, tightly integrated client use to actually replicate files from site to site answers question How do we move the files? * Slide courtesy of Scott Koranda, UW-Milwaukee

15. eSS 2007Service-Oriented Science: Globus Software in Action15 LDR Architecture l Each site has its own machinery for pulling data needed to satisfy local users l Scheduler daemon queries metadata and replica catalogs to identify missing files, which it records in a prioritized list l Transfer daemon manages transfers for files on the list, and updates LRC when files arrive l If a transfer fails, the transfer daemon simply re-adds the file to the list for future transfers l Daemons and metadata service written in Python

16. eSS 2007Service-Oriented Science: Globus Software in Action16 Results - LIGO * l LIGO/GEO S4 science run completed March u Replicated over 30 TB in 30 days u Mean time between failure now one month (outside of CIT) u Over 30 million files in LDR network l 6+ million unique LFNs in RLS network l 30+ million PFNs for those LFNs l Performance currently appears to be limited by Python performance l Partnership with Globus Alliance members is leading to LDR-like tools being added to Globus Toolkit * Slide courtesy of Scott Koranda, UW-Milwaukee

17. eSS 2007Service-Oriented Science: Globus Software in Action17 Results - GT and e-Science l Data Replication Service (DRS) u Reimplementation of LDRs subscription capabilities u Uses WSRF-compliant services written in Java u Works with RLS and RFT services in GT4 l Tech preview in GT4 l LIGO evaluating as a replacement for LDRs subscription system l Remaining piece is metadata service

18. eSS 2007Service-Oriented Science: Globus Software in Action18 A Few LDR/DRS Lessons l Understanding performance requirements is important u Performance is complicated. Requirements must be expressed very specifically and must be thoroughly informed by use u LIGOs operational use was specific enough to give the RLS team good requirements l Reliability doesnt have to be sophisticated u LDRs strategy for recovering from a transfer failure is simply to put the transfer request back on the queue l Standard interfaces are important u LIGO was not successful at getting other projects to rally around LDRs non-standard interfaces u DRS might overcome that hurdle by using WSRF interfaces